On this page
An ordinary-language question can remove a great deal of complexity from analytical software. But it still asks something of the person using it: they need to have a question. I found that this is a more significant limitation than it first appears.
The blank question is a design decision
A familiar approach to conversational analytics is to put a text box in front of a dataset. It makes the interface look simple. A person can apparently ask anything, with none of the measures, dimensions, filters or chart controls found in conventional business intelligence software.
The difficulty is that the blank box conceals a considerable amount of work. The user must decide what is worth investigating, recognise which questions the available data can support, choose the appropriate level of specificity and phrase the request in a way the system can understand. An experienced analyst may do much of that almost unconsciously. Someone responsible for a business process, rather than for analysing its data, may not.
I wanted to distinguish not knowing the answer from not knowing which question to ask. They are different problems and deserve different starting points.
Different entrances, one calculation authority
In determinIQ, Ask is intended for someone who can express a question in ordinary language. Explore is intended for someone who would benefit from a guided route through the dimensions and measures the dataset can support. Summary offers an initial orientation without requiring a question at all. Insights presents findings that have passed the system’s governed qualification process. Change-oriented routes help a person examine what has moved between comparable states.
These are not meant to be competing analytical engines. Each is an approach to a shared, governed analytical process. The distinction is important: if each interface could invent its own calculation rules, apparently equivalent questions could produce different answers merely because the user entered through a different screen.
Research figure 01
Several ways into the same analysis
Several routes to an answer, with one calculation authority
- BUSINESS NEEDUnderstand the available dataA person arrives with a question or a need for orientation
- CHOICE OF ENTRYAsk · Explore · Summary · InsightsChoose the route that suits the immediate need
- GOVERNED PREPARATIONEstablish a supported requestThe selected route cannot bypass validity or constraints
- ANALYTICAL AUTHORITYCalculate deterministicallyOne governed engine establishes permitted results
- EVIDENCE & FOLLOW-UPInspect and investigate furtherThe person can review the result or ask another question
A conceptual decision path, not a required sequence of user screens.
An interface can therefore be inviting without relaxing the boundary between intention and execution. The user may choose a guided analytical route, but the question still needs to be prepared and validated before a supported calculation takes place.
Guidance should come from the data, not guesswork
Explore is more useful when the routes it offers are relevant to the uploaded dataset. A time-based movement analysis is not meaningful without a usable time field. A comparison requires a compatible basis for comparison. A measure and a grouping field must have suitable analytical roles. The system should not offer an attractive action simply because it has a chart template available.
The documented Ask and Explore work has therefore involved classification, supported operator choices, field meaning and preparation-state controls. The goal is to translate the structure and capabilities of the data into a small set of understandable investigations, not to expose every internal operation as another option.
This is one of the less visible aspects of accessibility. Providing hundreds of controls and letting the user determine which are valid is not necessarily empowering. Sometimes the more useful design decision is to do the admissibility work in the system and present only meaningful choices.
The interface must preserve the user’s place
Guided exploration only helps if the system remains understandable after the first click. Users move between preparing a question, resolving an ambiguity, inspecting a result, following an explanation and returning to investigate something else. Each transition introduces the possibility of lost context.
A question that has already been answered should not silently become an unprepared draft. A user returning from an evidence view should be able to recognise the analysis they were inspecting. A clarification should relate to the pending question, rather than appearing as an unrelated conversation. These details may look like interface polish, but they are part of whether the analytical system can be used reliably.
Much of my later work on determinIQ has involved this kind of convergence: simplifying navigation, removing duplicate controls, clarifying selection states and preserving useful continuity. The objective is not simply a smaller interface. It is a system in which the person can understand what happened, what remains active and what they can do next.
Not all help is an AI suggestion
There is a temptation to equate guidance with a model proposing questions. That can be useful in a bounded role, but it is not sufficient. A suggested question still has to be supportable by the actual dataset and by the implemented analytical operators.
I prefer a separation between the system’s knowledge of its permitted analytical routes and any language assistance used to describe them. The system should be able to refuse unsupported requests and explain why a route is not available. A graceful refusal is often more useful than a question suggestion that ends in an invented result.
The same principle applies to automated discovery. A finding presented by Insights must come through the governed qualification path; it is not made significant by the enthusiasm of its explanation.
What this work has demonstrated
The current implementation includes distinct Summary, Ask, Explore and Insights experiences, with documented work on preparation, supported operators, navigation continuity and the governed discovery hand-off. The October 2026 research paper records why these complementary entrances were introduced and how they relate to the common analytical process.
That is an implementation and design account, not proof that every possible unfamiliar dataset will be understood correctly or that guided exploration eliminates all learning effort. Dataset semantics, missing values, ambiguous measures and unsupported questions remain practical limits. These limits are useful test cases precisely because the purpose of guided exploration is to expose supported possibilities rather than conceal unsupported ones.
The research question is whether we can make useful analytical work easier to begin without lowering the standard required to complete it. My working answer is to keep the entrances flexible, the execution authority singular and the user’s sense of continuity intact.
Connected work
This study sits alongside the research on analytical authority and evidence. The governed question contract matters whether a question began in Ask or Explore. Evidence matters once a result has been reached. A related journal entry considers a less obvious consequence: when the evidence does not justify a notable finding, returning no Insights is a valid outcome.
Research context
Research programme: Making data analysis understandable, reproducible and accountable
Topic: Governed analytics
Explore further
Supporting material is listed below.
Further reading is listed below.